If AI replaces musicians, does the entire plugin industry die with them?

[marcemad10]

I think they will separate "musical ai" subscription from plugin industry, you will probably have to pay for each of them.

[vurt]

(if that friend had a hidden agenda, and wasn't actually your friend)

a new canary in yellow by the light switch, who watches over you?

[vurt]

Wow, really?, I spent about 40 USD for music since 2004, when I bought my last CDs.
If I was 20 in 80ies or 70ies now, I would probably by one or two albums every week.
Let us say an album has 10 songs (average) and 80 albums per year. 800 songs to listen to in one year, makes more sense than having to find 800 nice songs amongst more than 50 Mio. new song/track releases every year nowadays.

But the way to find new music hasn't changed though. There's definitely more ways of doing it, but even then they are just variations on an existing theme.

Radio, magazines, concerts, friends all still exist. They might exist in different ways now, but that is just modernisation.

Even those algorithmically suggested songs are a bit like a friend suggesting something to you (if that friend had a hidden agenda, and wasn't actually your friend)

speaking of, the algorithm fed me dream eaters, so it's not all evil. unless you think satanism is evil of course....

[BBFG#]

Pondering when an AI manager to manage all the implemented AIs will be the next thing. Or if these type of threads will be along the lines of "Stock AI vs. Third Party AI", etc.
Oh so many rabbit holes...

[Zeisner]

Imagine not being able to upload tracks anymore, only prompts or seeds...

[wagtunes]

And the sky continues to fall.

[BONES]

"prompt creative ideas" With the LLM you'll be barking up the wrong tree. It doesn't have ideas.

Of course it has ideas. I just asked Co-Pilot for some - literally "give us some ideas" -

Those are some person's ideas, fuckwit.

Go and find the f**king things, then. Because I looked and couldn't find anything even remotely similar, despite multiple searches using different queries.

The willful ignorance on display here is staggering. You're sticking your fingers in your ears and singing "la-la-la" at the top of your voice like a f**king child. Grow the f**k up.

[BONES]

Wow, really?, I spent about 40 USD for music since 2004, when I bought my last CDs.
If I was 20 in 80ies or 70ies now, I would probably by one or two albums every week.
Let us say an album has 10 songs (average) and 80 albums per year. 800 songs to listen to in one year, makes more sense than having to find 800 nice songs amongst more than 50 Mio. new song/track releases every year nowadays. 50 Mio!!!!!!!!!!!!!
This is insane and it will be a lot more in future, maybe 50 Mio per month.
Only one new musician since 2004 that impressed me, I want to be impressed by lyrics, vibe, arrangement, performance, originality, voice, maybe virtuosity not just by average garbage.
I still find great music or performances from the padt, though. Charles Valentine Alkan was a cool composer and pianist in 19th century.

Again, I find those criteria to be largely superficial. It's the way a piece of music affects me, the visceral reaction it provokes. It's hard to put into words but it is to do with the "energy" and "attitude" of the piece that provokes a sort of cathartic response in me. When I used to have a really stressful day, I didn't come home and put something soothing on, I'd put on some Killing Joke, turn it up as loud as it would go and thrash around to it until I felt better. Getting up on stage provides that same catharsis, in spades. I get to jump around on stage and yell my lungs out for an hour and it feels f**king great. But it's all about "feel", the emotional response to the music. I find I often prefer to listen to a live album, where there is more of that raw energy, which often gets smoothed over in the studio.

When I was 20-21, doing officer training, I used to drive the best part of an hour into "town" every pay day (we got paid fortnightly), make a bee-line to the record shop and buy an LP from their "Punk/New Wave/Alternative bin. Sometimes I'd listen to it first but usually I'd pick something based on the cover artwork or the song names. I rarely bought a dud and there are albums I bought during that time that I still listen to regularly, 45 years later. Once you know what you're looking for, it's fairly easy to find the right stuff.

It would be much harder for you, because you'd almost certainly need to hear it first to know if it will meet your exacting criteria or not.

Anyway, this is all a bit off-topic. Maybe it could be useful to start a new thread where people can share their methods of discovering new music?

[ghettosynth]

"prompt creative ideas" With the LLM you'll be barking up the wrong tree. It doesn't have ideas.

Of course it has ideas. I just asked Co-Pilot for some - literally "give us some ideas" -

Those are some person's ideas, fuckwit.

Go and find the f**king things, then. Because I looked and couldn't find anything even remotely similar, despite multiple searches using different queries.

You won't. The entire exchange represents a misunderstanding of how LLMs work and how they capture training data in weights. LLMs do produce novel output, that is, output that is not in the training corpus. Yes, they predict the next token, but that process is commonly misunderstood. LLMs absorb structure from the training data at multiple levels of scale. At a small scale, they learn spelling, grammar, and common word pairings. At a larger scale, they learn how sentences form arguments, how paragraphs develop ideas, and how different genres are organized. At the largest scale, they capture patterns of style, tone, and narrative flow across many examples. Rather than storing documents like a database, the model compresses recurring patterns across its training data into a distributed representation. When it generates text, it recombines these learned patterns in new ways, which is why the output is usually novel even though it reflects general structures seen during training.

So, it's not correct to think that LLMs can't create novel outputs from the latent representations in their training data. Moreover, it is more anomalous for them to create exact duplicates of existing work. When recognizable training text is reproduced, it typically happens under highly constraining or adversarial prompts. If a user provides the beginning of a specific passage or strongly narrows the possible continuation, the model’s probability distribution collapses around the most statistically likely sequence, which may correspond to memorized text. In open ended settings, many different continuations are plausible, so the model samples more freely and produces original combinations of learned structures. Direct reproduction therefore tends to arise not from normal use, but from prompts designed to steer the model toward a particular known sequence.

Interestingly, the ability to produce coherent detailed facts of a specific domain is almost a side effect of the training on, e.g., vast numbers of papers and textbooks from that domain. Models learn patterns of language first and that domain specific factual output follows from modeling those patterns.

Interpolation between learned patterns drives originality in LLMs. They can recombine structures in ways that never appeared verbatim and extend them into new contexts, in text or music. Because their representations are continuous and compositional, the space of plausible outputs is much larger than the specific examples seen during training.

LLMs can create new work or what feels like new genres in much the same way humans do. Drum and Bass did not appear from nowhere. It recombined earlier rhythmic and electronic traditions into something distinct. What sounds new from a model usually comes from the same process, blending harmonic, rhythmic, and stylistic structures it has already internalized.

When researchers say LLMs cannot create outside their training data, they mean this in a narrow sense. A model trained only on Western harmony has no statistical basis for unfamiliar tonal systems. It can rearrange what it knows, but it cannot generate a coherent structure built on rules absent from its training distribution. In that strict sense, it does not create from nothing.

[BONES]

"AI is thinking, but it is also not conscious"
A contradiction in terms. It may tell you it's thinking but it doesn't have the first clue what that is.

You are confusing conscious and self-awareness with thought. Do you think when a leopard is stalking its prey that it's aware that it is thinking about all the factors that go into a successful hunt? Am I downwind? Is the grass long enough to hide me until I'm close enough to pounce? etc.? It is thinking about all those things but it's not doing it consciously. It doesn't need to know that it's thinking in order to think. The same is true of AI. It's responses may be programmed but it still needs to think about which is the appropriate response for any given situation. When a coach is selecting his team, he has criteria he must use to make the best selections. His thought process is algorithmic and constrained by certain criteria. AI is no different.

Intelligence is a part of consciousness.

Yes but consciousness is not a requirement for intelligence. Consciousness is built, to a degree, upon intelligence but intelligence exists all around us without consciousness. What your suggesting is akin to saying that cars need an engine, so if you don't have a car, you can't have an engine. It's more nonsense.

Consciousness is not computational - Roger Penrose.

You're the only one confusing thought and consciousness, Penrose would know better. This is a perfect example of the Appeal to Authority Fallacy you incorrectly accused me of in the other thread. So well done, you.

Don't listen to me:

---

At 7:40 he says "... and that's where it (ChatGPT) acquires its intelligence from". So he's acknowledging that ChatGPT has intelligence.
At 9:33 he says "so it's a fundamentally different kind of intelligence", which is exactly what I've been telling you for weeks. I'll listen to the rest of it later, I don't have time now.

Anything I say is derived strictly from my understanding from those with expertise and I'm not hallucinating.

Perhaps not but you are clearly being very selective as to which bits you pay attention to and which you willfully ignore.

Art without expression is called kitsch. Any music AI creates is kitsch by definition.

Only if the AI is working without human direction, which is something that, at this time, it is not capable of. You, the user, inject the expression and the art, the AI is just your palette and brush. It's palette would fill several rooms, which is what makes it such a potentially useful tool.

[BONES]

Radio, magazines, concerts, friends all still exist. They might exist in different ways now, but that is just modernisation.

No, it goes far deeper than that. You can't honestly suggest that listening to the radio today will expose you to the same breadth and depth of music that you could hear on even the most commercial stations in the 80s. Sure, it might be out there somewhere on some spotty kid's podcast or internet radio but good f**king luck finding it. You only have to look at the charts today to see the truth of this. Almost all of the music I am into today is stuff I first heard on mainstream commercial radio in the 70s and 80s.

Of course, getting suggestions from friends only works if you know people with similar taste in music to you. Apart from my bandmate, I don't know anyone who has even remotely similar musical taste to me and most of the flow between him and me goes in the other direction.

Even those algorithmically suggested songs are a bit like a friend suggesting something to you (if that friend had a hidden agenda, and wasn't actually your friend)

I find some work a lot better than others. Zune's "Related Artists" selections were usually spot-on, whereas I find the hit rate on Bandcamp and Discogs to be very low.

Pondering when an AI manager to manage all the implemented AIs will be the next thing. Or if these type of threads will be along the lines of "Stock AI vs. Third Party AI", etc.

By a strange coincidence, I saw this ad on the telly just last night -

Imagine not being able to upload tracks anymore, only prompts or seeds...

I'd be fine with that. After all, songs are for performing, not uploading.

[Zeisner]

And once again:

https://en.wikipedia.org/wiki/Clustering_illusion

https://en.wikipedia.org/wiki/ELIZA_effect

And of course:

https://en.wikipedia.org/wiki/Intelligent_design

[Zeisner]

https://www.linkedin.com/pulse/stochast ... garg-saisf

The same applies to everything else, including anything related to sound. Example: Tell AI to use ILD+ITD with a ratio of 4:5 to pan the instruments of an orchestra. It can't do it and it will never be able to. Like it will never be able to create new genres or styles. That's way too complex.

[_leras]

No matter how well the robot can sing and play guitar, people will still throw drinks at it if it trys to perform at the pub.

[_leras]

Radio, magazines, concerts, friends all still exist. They might exist in different ways now, but that is just modernisation.

No, it goes far deeper than that. You can't honestly suggest that listening to the radio today will expose you to the same breadth and depth of music that you could hear on even the most commercial stations in the 80s. Sure, it might be out there somewhere on some spotty kid's podcast or internet radio but good f**king luck finding it.

You might be surprised at the range of music that gets added to reels, shorts and tik toks.

No reference to where the music came from or the context it was created in... But it's out there.